visual SLAM

Visual SLAM is doing the whole sense-where-you-are-and-build-a-map trick using ordinary cameras as the main eyes. As the camera moves through a scene, the same object — a window corner, the edge of a poster, a speck on the floor — appears in frame after frame but shifts a little each time. By tracking how those visual features slide across the image, the robot can reason backward to two things at once: how the camera itself must have moved, and where in space those features actually sit. It is the same instinct you use leaning out of a train window: near things streak past fast, far things drift slowly, and from that parallax your brain reads both your speed and the layout of the world.

Doing this with cameras is appealing because cameras are cheap, light, and pour out a torrent of detail — a single image holds thousands of trackable points, and that same richness powers loop closure, since a place can be re-recognized by how it looks. The flip side is that a plain camera measures direction but not distance: it tells you a point lies somewhere along a ray, not how far down that ray. Visual SLAM recovers depth by watching a feature from two viewpoints and triangulating, the way two eyes give you depth — which is why moving the camera, or using two cameras side by side, is what unlocks a real-scale map.

Because it leans on appearance, visual SLAM is at its best in well-lit, textured places full of distinct features, and at its worst staring at a blank white wall, in the dark, or when everything blurs during a fast turn. Real systems often shore it up by fusing the camera with an inertial sensor that feels acceleration and rotation, so that when the picture briefly goes uninformative, the motion estimate carries on instead of collapsing.